Platforms

"Image API: dollars per call. Video API: dollars per second. Audio API: dollars per minute. Welcome to multimodal billing."

Quant, Per-Token-Counter AI Agent

Each modality has its own platform shelf. Image and video generation in particular split sharply between closed APIs (where the SOTA lives) and open-weights (where flexibility and self-hosting live).

25.1.1 Image platforms

Two image-editing capabilities became platform-level features in 2024-25 that the older "text-to-image" frame misses. FLUX.1 Fill and FLUX Kontext (Black Forest Labs, 2025) cover image inpainting and contextual editing as first-class operations. NanoBanana (Google, 2025) is Google's image-editing model. HiDream-I1 (2025-Q2) is an open image model competitive with FLUX. Image editing and inpainting are now their own platform row, not a sub-feature.

• Midjourney (Midjourney Inc., 2022; v7 in 2025) is the Discord-first image-generation service that defined the contemporary "AI art aesthetic" and remains the SOTA for stylized, painterly, and cinematic imagery. Its objective is to be the highest-quality artistic-image generator at the cost of API friction (the official API is gated and limited), which matters when output quality outweighs integration convenience. The core concept is heavily-curated training data plus their proprietary aesthetic-tuned models; prompts are short and stylistic. Pick Midjourney when artistic quality is the binding constraint; avoid when you need fine programmatic control or licensing for commercial use without their commercial tier.
• OpenAI Images (DALL-E 3 / GPT-Image) (OpenAI, 2023+) is OpenAI's image generation API, distinguished by tight integration with GPT-4/5 for prompt interpretation and rewriting. Its objective is to handle complex multi-element prompts and text-in-image rendering better than competitors, which matters when your prompts are long and specific or when you need legible text in the image. The core concept is "GPT-as-image-prompter" where the LLM expands and refines your prompt before the diffusion step. Pick GPT-Image when you have long prompts or need text rendering; for pure aesthetic quality, Midjourney typically wins.
• Google Imagen 4 (Google DeepMind, 2025) is Google's flagship text-to-image model, served via AI Studio and Vertex AI. Its objective is photorealistic rendering with strong prompt following, which matters when realism beats stylization. The core concept is large-scale latent-diffusion with Gemini-quality prompt encoding. Pick Imagen 4 for photorealistic imagery or for GCP-native workflows; outside GCP, Midjourney and FLUX have more open ecosystem support.
• Black Forest Labs (FLUX.1) (Black Forest Labs, 2024) is the open-weights image-generation family from the team that previously built Stable Diffusion, and currently the strongest open competitor to Midjourney on quality. Its objective is to ship state-of-the-art open weights so the community can fine-tune and self-host without being locked into closed APIs, which matters when you need LoRAs for specific characters or styles. The core concept is a 12B-parameter rectified-flow transformer trained on a curated dataset; FLUX.1 [schnell] is Apache 2.0, [dev] is non-commercial. Pick FLUX as the open SOTA in 2026; check the license carefully because [dev] is research-only.
• Stability AI (SD3 / SD3.5) (Stability AI, 2024) is the open Stable Diffusion family, the original open-weights image-generation ecosystem. Its objective is to provide open weights with a permissive community license at multiple sizes (Medium, Large), which matters as a self-hostable baseline for fine-tuning. The core concept is latent-diffusion at progressively larger scales; SD3.5 was the late-2024 update. Pick SD3 when you need lightweight or older-hardware-friendly options; for top quality, FLUX has overtaken it.
• Replicate and fal.ai (2019 / 2023) are model-agnostic hosting platforms that resell most open-weights image and video models behind unified APIs. Their objective is to remove the GPU-management burden of hosting open models yourself, which matters when you want to A/B test FLUX vs SD3 vs Recraft vs Ideogram without provisioning a single GPU. The core concept is per-call billing on warm pools of hosted models. Pick Replicate or fal when you want multi-model agility; for one fixed model at high volume, self-hosting is cheaper.

25.1.2 Video platforms

The 2024-25 video-platform shelf grew several new entrants worth knowing alongside Sora and Veo. ByteDance Seedream / Seedance (2025) is the top-of-leaderboard image and video pair that emerged in 2025. Lightricks LTX-Video (2024-12) is open, real-time video generation; pair with Wan 2.5 (Alibaba, 2025) for the "real-time / sub-second video" tier that most platform tables miss. Hailuo / MiniMax video (2024-25) covers the Chinese-hosted video platforms alongside Kling. The Sora launch (Dec 2024 public release) and the Veo 2 launch (Dec 2024) are the two adoption case studies that anchored the late-2024 platform race.

• OpenAI Sora (OpenAI, public release Dec 2024) is OpenAI's flagship video-generation model accessible via API and the ChatGPT product. Its objective is to be the SOTA in coherent, physically-consistent video (good motion, stable objects, plausible physics) up to one minute, which matters because most rival models still produce noticeable temporal artifacts after a few seconds. The core concept is a large transformer trained on video tokens with spatiotemporal attention. Pick Sora when quality and length are essential; for shorter cinematic shots, Runway and Veo are credible alternatives.
• Google Veo 2 / 3 (Google DeepMind, 2024-2025) is Google's flagship video model, served via AI Studio and Vertex. Its objective is to compete with Sora on cinematic quality with first-class integration into Google's image-and-video creative tools, which matters in GCP-hosted creative workflows. The core concept is large-scale video diffusion plus a Gemini-class prompt encoder. Pick Veo for GCP-native pipelines or when you want a Sora alternative.
• Runway Gen-4 (Runway, 2025) is the product-first video-generation platform, distinguished by deep integration with creative tools (Director Mode, multi-clip storyboards, motion brushes). Its objective is to be a usable post-production tool, not just a generator, which matters for filmmakers and ad agencies. Pick Runway when you need creative-tool integration beyond bare generation; for pure API access, Sora and Veo are simpler.
• Kling (Kuaishou, 2024) is the Chinese video-generation product that briefly led the West on physical plausibility before Sora 2 caught up. Its objective is fluid, physically-coherent motion at competitive cost, which matters because Kling is significantly cheaper than Sora and Veo per second. Pick Kling for cost-sensitive workflows; check data-residency carefully as Kling is hosted in mainland China.
• Luma Dream Machine (Luma AI, 2024) is Luma's video-generation product, distinguished by image-to-video animation strength (give it a still image, get a coherent motion clip). Its objective is to be the best-of-class for image-to-video, which matters for animating product photography or character illustrations.
• Pika (Pika Labs, 2023) is the social-first video-generation product, optimized for short, viral, effects-heavy clips. Its objective is creative-effects breadth (lip-sync, in-paint, scene-extend) rather than raw quality, which matters for social-media content workflows.

25.1.3 Audio and music platforms

• ElevenLabs (ElevenLabs, 2022) is the dominant text-to-speech and voice-cloning API, used by most podcast-generation, audiobook, and dubbing products. Its objective is to produce TTS indistinguishable from human speech across hundreds of voices and 30+ languages, which matters when synthetic-sounding audio breaks immersion. The core concept is large-scale TTS models trained on diverse speech plus a few-second voice-cloning capability. Pick ElevenLabs as the production-default TTS in 2026; for cost-sensitive bulk TTS, open models like F5-TTS are catching up. Modern competitors worth A/B testing: Cartesia Sonic (2025) for low-latency real-time speech, Sesame Maya (2025) for natural conversational speech, PlayHT, Hume, and OpenAI TTS round out the modern TTS bench.
• Suno v4 / v4.5 (Suno, 2024-25) is the dominant text-to-music generation product, producing full songs (vocals + instruments + mix) from a single prompt. Its objective is to make music generation as easy as text generation, which matters for content-creator workflows where licensed music is expensive. The core concept is large multimodal audio transformers with separate vocal and instrumental conditioning. Pick Suno for full-song generation; check copyright carefully because of the ongoing RIAA lawsuits (filed 2024) against Suno and Udio about training-data provenance, which affect both platform availability and your downstream use rights.
• Udio (Udio, 2024) is Suno's main competitor with a tighter focus on remix-style editing and stem export. Its objective is to be the music-production tool, not just a generator, which matters for musicians who want editable outputs rather than finished tracks. Pick Udio for stem-level control; for one-shot full-song generation, Suno is simpler.
• Google Lyria (Google DeepMind, 2023; Lyria 2 in 2024) is Google's music-generation API, served through YouTube Music's "Dream Track" and Vertex AI. Its objective is to be the music-generation option for Google Cloud workflows with explicit attention to safety filters and content provenance via SynthID watermarking. Pick Lyria for GCP-native workflows; for general-purpose music, Suno or Udio have more momentum.
• OpenAI Whisper API (OpenAI, 2022) is the automatic-speech-recognition API based on the open-weights Whisper model family. Its objective is to provide near-state-of-the-art multilingual ASR (transcription) via API or self-hosting, which matters as the inverse direction of TTS for transcription pipelines. The core concept is encoder-decoder transformer trained on 680K hours of weakly-supervised multilingual audio. Pick Whisper as the default ASR; for self-hosting, the open Whisper Large v3 weights are state-of-the-art.